Parametrizing tests

pytest allows to easily parametrize test functions. For basic docs, see 如何参数化 fixtures 和测试函数.

In the following we provide some examples using the builtin mechanisms.

Generating parameters combinations, depending on command line

Let’s say we want to execute a test with different computation parameters and the parameter range shall be determined by a command line argument. Let’s first write a simple (do-nothing) computation test:

# content of test_compute.py


def test_compute(param1):
    assert param1 < 4

Now we add a test configuration like this:

# content of conftest.py


def pytest_addoption(parser):
    parser.addoption("--all", action="store_true", help="run all combinations")


def pytest_generate_tests(metafunc):
    if "param1" in metafunc.fixturenames:
        if metafunc.config.getoption("all"):
            end = 5
        else:
            end = 2
        metafunc.parametrize("param1", range(end))

This means that we only run 2 tests if we do not pass --all:

$ pytest -q test_compute.py
..                                                                   [100%]
2 passed in 0.12s

We run only two computations, so we see two dots. let’s run the full monty:

$ pytest -q --all
....F                                                                [100%]
================================= FAILURES =================================
_____________________________ test_compute[4] ______________________________

param1 = 4

    def test_compute(param1):
>       assert param1 < 4
E       assert 4 < 4

test_compute.py:4: AssertionError
========================= short test summary info ==========================
FAILED test_compute.py::test_compute[4] - assert 4 < 4
1 failed, 4 passed in 0.12s

As expected when running the full range of param1 values we’ll get an error on the last one.

Different options for test IDs

pytest will build a string that is the test ID for each set of values in a parametrized test. These IDs can be used with -k to select specific cases to run, and they will also identify the specific case when one is failing. Running pytest with --collect-only will show the generated IDs.

Numbers, strings, booleans and None will have their usual string representation used in the test ID. For other objects, pytest will make a string based on the argument name:

# content of test_time.py

from datetime import datetime, timedelta

import pytest

testdata = [
    (datetime(2001, 12, 12), datetime(2001, 12, 11), timedelta(1)),
    (datetime(2001, 12, 11), datetime(2001, 12, 12), timedelta(-1)),
]


@pytest.mark.parametrize("a,b,expected", testdata)
def test_timedistance_v0(a, b, expected):
    diff = a - b
    assert diff == expected


@pytest.mark.parametrize("a,b,expected", testdata, ids=["forward", "backward"])
def test_timedistance_v1(a, b, expected):
    diff = a - b
    assert diff == expected


def idfn(val):
    if isinstance(val, (datetime,)):
        # note this wouldn't show any hours/minutes/seconds
        return val.strftime("%Y%m%d")


@pytest.mark.parametrize("a,b,expected", testdata, ids=idfn)
def test_timedistance_v2(a, b, expected):
    diff = a - b
    assert diff == expected


@pytest.mark.parametrize(
    "a,b,expected",
    [
        pytest.param(
            datetime(2001, 12, 12), datetime(2001, 12, 11), timedelta(1), id="forward"
        ),
        pytest.param(
            datetime(2001, 12, 11), datetime(2001, 12, 12), timedelta(-1), id="backward"
        ),
    ],
)
def test_timedistance_v3(a, b, expected):
    diff = a - b
    assert diff == expected

In test_timedistance_v0, we let pytest generate the test IDs.

In test_timedistance_v1, we specified ids as a list of strings which were used as the test IDs. These are succinct, but can be a pain to maintain.

In test_timedistance_v2, we specified ids as a function that can generate a string representation to make part of the test ID. So our datetime values use the label generated by idfn, but because we didn’t generate a label for timedelta objects, they are still using the default pytest representation:

$ pytest test_time.py --collect-only
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y
rootdir: /home/sweet/project
collected 8 items

<Module test_time.py>
  <Function test_timedistance_v0[a0-b0-expected0]>
  <Function test_timedistance_v0[a1-b1-expected1]>
  <Function test_timedistance_v1[forward]>
  <Function test_timedistance_v1[backward]>
  <Function test_timedistance_v2[20011212-20011211-expected0]>
  <Function test_timedistance_v2[20011211-20011212-expected1]>
  <Function test_timedistance_v3[forward]>
  <Function test_timedistance_v3[backward]>

======================== 8 tests collected in 0.12s ========================

In test_timedistance_v3, we used pytest.param to specify the test IDs together with the actual data, instead of listing them separately.

A quick port of “testscenarios”

Here is a quick port to run tests configured with testscenarios, an add-on from Robert Collins for the standard unittest framework. We only have to work a bit to construct the correct arguments for pytest’s Metafunc.parametrize():

# content of test_scenarios.py


def pytest_generate_tests(metafunc):
    idlist = []
    argvalues = []
    for scenario in metafunc.cls.scenarios:
        idlist.append(scenario[0])
        items = scenario[1].items()
        argnames = [x[0] for x in items]
        argvalues.append([x[1] for x in items])
    metafunc.parametrize(argnames, argvalues, ids=idlist, scope="class")


scenario1 = ("basic", {"attribute": "value"})
scenario2 = ("advanced", {"attribute": "value2"})


class TestSampleWithScenarios:
    scenarios = [scenario1, scenario2]

    def test_demo1(self, attribute):
        assert isinstance(attribute, str)

    def test_demo2(self, attribute):
        assert isinstance(attribute, str)

this is a fully self-contained example which you can run with:

$ pytest test_scenarios.py
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y
rootdir: /home/sweet/project
collected 4 items

test_scenarios.py ....                                               [100%]

============================ 4 passed in 0.12s =============================

If you just collect tests you’ll also nicely see ‘advanced’ and ‘basic’ as variants for the test function:

$ pytest --collect-only test_scenarios.py
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y
rootdir: /home/sweet/project
collected 4 items

<Module test_scenarios.py>
  <Class TestSampleWithScenarios>
    <Function test_demo1[basic]>
    <Function test_demo2[basic]>
    <Function test_demo1[advanced]>
    <Function test_demo2[advanced]>

======================== 4 tests collected in 0.12s ========================

Note that we told metafunc.parametrize() that your scenario values should be considered class-scoped. With pytest-2.3 this leads to a resource-based ordering.

Deferring the setup of parametrized resources

The parametrization of test functions happens at collection time. It is a good idea to setup expensive resources like DB connections or subprocess only when the actual test is run. Here is a simple example how you can achieve that. This test requires a db object fixture:

# content of test_backends.py

import pytest


def test_db_initialized(db):
    # a dummy test
    if db.__class__.__name__ == "DB2":
        pytest.fail("deliberately failing for demo purposes")

We can now add a test configuration that generates two invocations of the test_db_initialized function and also implements a factory that creates a database object for the actual test invocations:

# content of conftest.py
import pytest


def pytest_generate_tests(metafunc):
    if "db" in metafunc.fixturenames:
        metafunc.parametrize("db", ["d1", "d2"], indirect=True)


class DB1:
    "one database object"


class DB2:
    "alternative database object"


@pytest.fixture
def db(request):
    if request.param == "d1":
        return DB1()
    elif request.param == "d2":
        return DB2()
    else:
        raise ValueError("invalid internal test config")

Let’s first see how it looks like at collection time:

$ pytest test_backends.py --collect-only
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y
rootdir: /home/sweet/project
collected 2 items

<Module test_backends.py>
  <Function test_db_initialized[d1]>
  <Function test_db_initialized[d2]>

======================== 2 tests collected in 0.12s ========================

And then when we run the test:

$ pytest -q test_backends.py
.F                                                                   [100%]
================================= FAILURES =================================
_________________________ test_db_initialized[d2] __________________________

db = <conftest.DB2 object at 0xdeadbeef0001>

    def test_db_initialized(db):
        # a dummy test
        if db.__class__.__name__ == "DB2":
>           pytest.fail("deliberately failing for demo purposes")
E           Failed: deliberately failing for demo purposes

test_backends.py:8: Failed
========================= short test summary info ==========================
FAILED test_backends.py::test_db_initialized[d2] - Failed: deliberately f...
1 failed, 1 passed in 0.12s

The first invocation with db == "DB1" passed while the second with db == "DB2" failed. Our db fixture function has instantiated each of the DB values during the setup phase while the pytest_generate_tests generated two according calls to the test_db_initialized during the collection phase.

Indirect parametrization

Using the indirect=True parameter when parametrizing a test allows to parametrize a test with a fixture receiving the values before passing them to a test:

import pytest


@pytest.fixture
def fixt(request):
    return request.param * 3


@pytest.mark.parametrize("fixt", ["a", "b"], indirect=True)
def test_indirect(fixt):
    assert len(fixt) == 3

This can be used, for example, to do more expensive setup at test run time in the fixture, rather than having to run those setup steps at collection time.

Apply indirect on particular arguments

Very often parametrization uses more than one argument name. There is opportunity to apply indirect parameter on particular arguments. It can be done by passing list or tuple of arguments’ names to indirect. In the example below there is a function test_indirect which uses two fixtures: x and y. Here we give to indirect the list, which contains the name of the fixture x. The indirect parameter will be applied to this argument only, and the value a will be passed to respective fixture function:

# content of test_indirect_list.py

import pytest


@pytest.fixture(scope="function")
def x(request):
    return request.param * 3


@pytest.fixture(scope="function")
def y(request):
    return request.param * 2


@pytest.mark.parametrize("x, y", [("a", "b")], indirect=["x"])
def test_indirect(x, y):
    assert x == "aaa"
    assert y == "b"

The result of this test will be successful:

$ pytest -v test_indirect_list.py
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y -- $PYTHON_PREFIX/bin/python
cachedir: .pytest_cache
rootdir: /home/sweet/project
collecting ... collected 1 item

test_indirect_list.py::test_indirect[a-b] PASSED                     [100%]

============================ 1 passed in 0.12s =============================

Parametrizing test methods through per-class configuration

Here is an example pytest_generate_tests function implementing a parametrization scheme similar to Michael Foord’s unittest parametrizer but in a lot less code:

# content of ./test_parametrize.py
import pytest


def pytest_generate_tests(metafunc):
    # called once per each test function
    funcarglist = metafunc.cls.params[metafunc.function.__name__]
    argnames = sorted(funcarglist[0])
    metafunc.parametrize(
        argnames, [[funcargs[name] for name in argnames] for funcargs in funcarglist]
    )


class TestClass:
    # a map specifying multiple argument sets for a test method
    params = {
        "test_equals": [dict(a=1, b=2), dict(a=3, b=3)],
        "test_zerodivision": [dict(a=1, b=0)],
    }

    def test_equals(self, a, b):
        assert a == b

    def test_zerodivision(self, a, b):
        with pytest.raises(ZeroDivisionError):
            a / b

Our test generator looks up a class-level definition which specifies which argument sets to use for each test function. Let’s run it:

$ pytest -q
F..                                                                  [100%]
================================= FAILURES =================================
________________________ TestClass.test_equals[1-2] ________________________

self = <test_parametrize.TestClass object at 0xdeadbeef0002>, a = 1, b = 2

    def test_equals(self, a, b):
>       assert a == b
E       assert 1 == 2

test_parametrize.py:21: AssertionError
========================= short test summary info ==========================
FAILED test_parametrize.py::TestClass::test_equals[1-2] - assert 1 == 2
1 failed, 2 passed in 0.12s

Indirect parametrization with multiple fixtures

Here is a stripped down real-life example of using parametrized testing for testing serialization of objects between different python interpreters. We define a test_basic_objects function which is to be run with different sets of arguments for its three arguments:

• python1: first python interpreter, run to pickle-dump an object to a file

• python2: second interpreter, run to pickle-load an object from a file

• obj: object to be dumped/loaded

"""
module containing a parametrized tests testing cross-python
serialization via the pickle module.
"""
import shutil
import subprocess
import textwrap

import pytest

pythonlist = ["python3.5", "python3.6", "python3.7"]


@pytest.fixture(params=pythonlist)
def python1(request, tmp_path):
    picklefile = tmp_path / "data.pickle"
    return Python(request.param, picklefile)


@pytest.fixture(params=pythonlist)
def python2(request, python1):
    return Python(request.param, python1.picklefile)


class Python:
    def __init__(self, version, picklefile):
        self.pythonpath = shutil.which(version)
        if not self.pythonpath:
            pytest.skip(f"{version!r} not found")
        self.picklefile = picklefile

    def dumps(self, obj):
        dumpfile = self.picklefile.with_name("dump.py")
        dumpfile.write_text(
            textwrap.dedent(
                r"""
                import pickle
                f = open({!r}, 'wb')
                s = pickle.dump({!r}, f, protocol=2)
                f.close()
                """.format(
                    str(self.picklefile), obj
                )
            )
        )
        subprocess.check_call((self.pythonpath, str(dumpfile)))

    def load_and_is_true(self, expression):
        loadfile = self.picklefile.with_name("load.py")
        loadfile.write_text(
            textwrap.dedent(
                r"""
                import pickle
                f = open({!r}, 'rb')
                obj = pickle.load(f)
                f.close()
                res = eval({!r})
                if not res:
                    raise SystemExit(1)
                """.format(
                    str(self.picklefile), expression
                )
            )
        )
        print(loadfile)
        subprocess.check_call((self.pythonpath, str(loadfile)))


@pytest.mark.parametrize("obj", [42, {}, {1: 3}])
def test_basic_objects(python1, python2, obj):
    python1.dumps(obj)
    python2.load_and_is_true(f"obj == {obj}")

Running it results in some skips if we don’t have all the python interpreters installed and otherwise runs all combinations (3 interpreters times 3 interpreters times 3 objects to serialize/deserialize):

. $ pytest -rs -q multipython.py
sssssssssssssssssssssssssss                                          [100%]
========================= short test summary info ==========================
SKIPPED [9] multipython.py:29: 'python3.5' not found
SKIPPED [9] multipython.py:29: 'python3.6' not found
SKIPPED [9] multipython.py:29: 'python3.7' not found
27 skipped in 0.12s

Indirect parametrization of optional implementations/imports

If you want to compare the outcomes of several implementations of a given API, you can write test functions that receive the already imported implementations and get skipped in case the implementation is not importable/available. Let’s say we have a “base” implementation and the other (possibly optimized ones) need to provide similar results:

# content of conftest.py

import pytest


@pytest.fixture(scope="session")
def basemod(request):
    return pytest.importorskip("base")


@pytest.fixture(scope="session", params=["opt1", "opt2"])
def optmod(request):
    return pytest.importorskip(request.param)

And then a base implementation of a simple function:

# content of base.py
def func1():
    return 1

And an optimized version:

# content of opt1.py
def func1():
    return 1.0001

And finally a little test module:

# content of test_module.py


def test_func1(basemod, optmod):
    assert round(basemod.func1(), 3) == round(optmod.func1(), 3)

If you run this with reporting for skips enabled:

$ pytest -rs test_module.py
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y
rootdir: /home/sweet/project
collected 2 items

test_module.py .s                                                    [100%]

========================= short test summary info ==========================
SKIPPED [1] conftest.py:12: could not import 'opt2': No module named 'opt2'
======================= 1 passed, 1 skipped in 0.12s =======================

You’ll see that we don’t have an opt2 module and thus the second test run of our test_func1 was skipped. A few notes:

• the fixture functions in the conftest.py file are “session-scoped” because we don’t need to import more than once

• if you have multiple test functions and a skipped import, you will see the [1] count increasing in the report

• you can put @pytest.mark.parametrize style parametrization on the test functions to parametrize input/output values as well.

Set marks or test ID for individual parametrized test

Use pytest.param to apply marks or set test ID to individual parametrized test. For example:

# content of test_pytest_param_example.py
import pytest


@pytest.mark.parametrize(
    "test_input,expected",
    [
        ("3+5", 8),
        pytest.param("1+7", 8, marks=pytest.mark.basic),
        pytest.param("2+4", 6, marks=pytest.mark.basic, id="basic_2+4"),
        pytest.param(
            "6*9", 42, marks=[pytest.mark.basic, pytest.mark.xfail], id="basic_6*9"
        ),
    ],
)
def test_eval(test_input, expected):
    assert eval(test_input) == expected

In this example, we have 4 parametrized tests. Except for the first test, we mark the rest three parametrized tests with the custom marker basic, and for the fourth test we also use the built-in mark xfail to indicate this test is expected to fail. For explicitness, we set test ids for some tests.

Then run pytest with verbose mode and with only the basic marker:

$ pytest -v -m basic
=========================== test session starts ============================
platform linux -- Python 3.x.y, pytest-7.x.y, pluggy-1.x.y -- $PYTHON_PREFIX/bin/python
cachedir: .pytest_cache
rootdir: /home/sweet/project
collecting ... collected 24 items / 21 deselected / 3 selected

test_pytest_param_example.py::test_eval[1+7-8] PASSED                [ 33%]
test_pytest_param_example.py::test_eval[basic_2+4] PASSED            [ 66%]
test_pytest_param_example.py::test_eval[basic_6*9] XFAIL             [100%]

=============== 2 passed, 21 deselected, 1 xfailed in 0.12s ================

As the result:

• Four tests were collected

• One test was deselected because it doesn’t have the basic mark.

• Three tests with the basic mark was selected.

• The test test_eval[1+7-8] passed, but the name is autogenerated and confusing.

• The test test_eval[basic_2+4] passed.

• The test test_eval[basic_6*9] was expected to fail and did fail.

Parametrizing conditional raising

Use pytest.raises() with the pytest.mark.parametrize decorator to write parametrized tests in which some tests raise exceptions and others do not.

It may be helpful to use nullcontext as a complement to raises.

For example:

from contextlib import nullcontext as does_not_raise

import pytest


@pytest.mark.parametrize(
    "example_input,expectation",
    [
        (3, does_not_raise()),
        (2, does_not_raise()),
        (1, does_not_raise()),
        (0, pytest.raises(ZeroDivisionError)),
    ],
)
def test_division(example_input, expectation):
    """Test how much I know division."""
    with expectation:
        assert (6 / example_input) is not None

In the example above, the first three test cases should run unexceptionally, while the fourth should raise ZeroDivisionError.